M PRA
Munich Personal RePEc Archive
Characterization of multidimensional
spatial models of elections with a valence
dimension
Yaron Azrieli
The Ohio State University
30. March 2009
Online at http://mpra.ub.uni-muenchen.de/14513/
MPRA Paper No. 14513, posted 8. April 2009 14:15 UTC
CHARACTERIZATION OF MULTIDIMENSIONAL SPATIAL MODELS
OF ELECTIONS WITH A VALENCE DIMENSION
YARON AZRIELI†
March 30, 2009
Preliminary draft
Abstract. Spatial models of political competition are typically based on two assumptions. One is that all the voters identically perceive the platforms of the candidates
and agree about their score on a “valence” dimension. The second is that each voter’s
preferences over policies are decreasing in the distance from that voter’s ideal point,
and that valence scores enter the utility function in an additively separable way.
The goal of this paper is to examine the restrictions that these two assumptions impose, starting from a more primitive (and observable) data. Specifically, we consider the
case where only the ideal point in the policy space and the ranking over candidates are
known for each voter. We provide necessary and sufficient conditions for this collection
of preference relations to be consistent with utility maximization as in the standard
models described above. That is, we characterize the case where there are policies
x1 , . . . , xm for the m candidates and numbers v1 , . . . , vm representing valence scores,
such that a voter with an ideal policy y ranks the candidates according to vi −||xi −y||2 .
Keywords: Elections, Spatial model, Valence, Euclidean preferences.
JEL Classification: D72
†
I am grateful to Ehud Lehrer for suggesting this problem. I would also like to thank Itzhak Gilboa,
Paul Healy, Dan Levin and Jim Peck for helpful comments and references.
Department of Economics, The Ohio State University, 1945 North High street, Columbus, OH 43210.
e-mail: [email protected].
1
2
YARON AZRIELI
1. Introduction
Since the seminal works of Hotelling [15] and Downs [10], spatial models of elections
are widely used in the political economy literature. Typically, these models identify the
policy space with a finite dimensional Euclidean space. Each voter in the electorate
is assumed to have an ideal point in the policy space. Candidates then choose their
platforms and each voter votes for the candidate with the closest platform to her ideal
policy. Usually, the emphasis is on equilibrium analysis of the resulting game between
candidates. This literature is too enormous to mention specific works.
More recently, authors incorporated a “valence” dimension to the standard model.
This additional dimension influence voters’ preferences and was shown to have a dramatic
effect on the outcome of the political game, both in theory and in empirical studies. This
additional dimension may represent any non–policy issue on which candidates differ in
the “score” they get from the electorate. Examples include charisma, experience, past
success, communication skills, etc. The reader can find references to many works that
discuss this point in the related literature section below.
When valence issues are present, the utility of a voter with an ideal point (in the policy
space) y if candidate i wins the elections is usually taken to be of the form vi − ||xi − y||2 ,
where vi and xi are the valence score and platform of candidate i respectively. This
particular form of utility function implies that each voter’s utility is decreasing in the
Euclidean distance between his own ideal point and the winning candidate platform, and
that valence scores enter the utility function in an additively separable way. Although
very natural, it is not clear why this particular functional form is more appropriate
than others to represent preferences. Possibly, using a different type of utility function
can make a difference for the predictions of a theoretical model and for the statistical
significance of an empirical model. Furthermore, an implicit assumption made here is
that all voters perceive in the same way each of the candidates’ platforms, and give the
same valence score to every candidate. That is, xi and vi are common to all the voters
for each i. This is certainly a crucial assumption since allowing voters to disagree on
candidates platforms and valence scores can lead to an intractable model.
Obviously, it is very hard (not to say impossible) to extract the entire preferences
of each voter over the policy space and the valence score he gives to each candidate.
Therefore, it is not easy to check whether the aforementioned assumptions make sense
CHARACTERIZATION OF SPATIAL MODELS
3
in any particular political campaign. Thus, an important matter is to identify conditions on more easily observable data that guarantee consistency with the spatial model
assumptions. Introducing such necessary and sufficient conditions is the main result of
this paper.
Specifically, we assume that, for each voter in the electorate, only the ideal policy and
the ranking of candidates can be observed. While this may also seem quite demanding,
it is much more reasonable than observing the entire utility function of the voter. We
characterize the case where this data is consistent with voters having utility functions
as above. That is, we characterize the case where there are policies x1 , . . . , xm for the m
candidates and numbers v1 , . . . , vm representing valence scores, such that a voter with
an ideal policy y ranks the candidates according to vi − ||xi − y||2 .
We use four conditions for the characterization. The first three are quite standard,
while the last one is more technical and is required for the proof. We think of our result
as “good news” since it shows that if the data satisfies rather weak assumptions then it
is consistent with the standard spatial model with a valence dimension1. However, our
model suffers from several weaknesses which make its applicability limited2. We therefore
think that the main contribution of this paper is to introduce the theoretical “revealed
preferences” approach to the literature on spatial models of elections. To the best of our
knowledge, this is the first attempt in this direction, and we hope it will lead to further
investigation.
1.1. Related literature. Papers using spatial models of elections with valence issues
similar to the one studied here are numerous in recent years. Examples include Ansolabehere and Snyder [1], Aragones and Palfrey [2], Degan [8], Dix and Santore [9],
Enelow and Hinich [11], Gersbach [12], Groseclose [14], Kim [16] and Schofield [19]
among others. These papers study different aspects of the political game and provide
various interpretations for the additive constant in the utility functions of the voters.
From a mathematical perspective, our main result is closely related to the result
in Azrieli and Lehrer [6], who characterize categorization systems that are generated
by proximity to a set of prototypical cases. Furthermore, there is a surprisingly close
connection between the result of this paper and the axiomatic derivation of Gilboa and
1Our
conditions are not sufficient if one doesn’t allow for a valence dimension, and we do not know
how to characterize data consistent with spatial models without this additive term. See subsection 3.4.
2See
Section 3.
4
YARON AZRIELI
Schmeidler [13] in their Case-Based Decision Theory3. The literature on scoring rules
(Myerson [18], Smith [20], Young [21]) is also closely related from a technical point of
view.
Finally, the mathematical object we deal with here is known in the geometry literature
as Voronoi diagram or Dirichlet tessellation. The most relevant papers in this literature
are Ash and Bolker [3], [4] and Aurenhammer [5]. The book by Boots et al. [7] surveys
applications of Voronoi diagrams in many different fields.
1.2. Organization. The next section contains the model and the main result of the
paper. In Section 3 we discuss in more detail several issues related to our model. In
particular, this section contains some additional results concerning the uniqueness of the
representation, the failure of our result if the set of voters is finite, and the special cases
of three candidates and one–dimensional policy space. We also point to some possible
directions of related future research. All the proofs are in Section 4.
2. Model and Result
Let C = {1, 2, . . . , m} be the set of candidates where m ≥ 2. The policy space is taken
to be Rd with4 d ≥ 2. Each potential voter is identified with her ideal point in the policy
space and we assume that, for every y ∈ Rd , there is a voter with y as her ideal policy.
Thus, the set of voters is also Rd . We will use the letters i, j, k, l to denote candidates
(elements of C) and x, y, z to denote voters or policies (points in Rd ).
Our primitive is a collection of binary relations {ºy }y∈Rd over C, one for every voter
y ∈ Rd . The interpretation of i ºy j is that a voter with an ideal point y (weakly)
prefers candidate i to candidate j. As usual, for any i, j ∈ C, we let i Ây j if and only if
both i ºy j and j y i, and i ∼y j if and only if both i ºy j and j ºy i. The following
properties will be used for the characterization.
(A1) Weak order: For every y ∈ Rd , ºy is complete and transitive.
(A2) Continuity: For every i, j ∈ C, the set {y ∈ Rd : i Ây j} is open.
(A3) Convexity: For every i, j ∈ C and y, z ∈ Rd , if i ºy j (i Ây j) and i ºz j then
i ºαy+(1−α)z j (i Âαy+(1−α)z j) for every α ∈ (0, 1).
3We
thank Itzhak Gilboa for pointing out this connection.
4Our
result does not hold in the case d = 1. We elaborate on this case in subsection 3.7.
CHARACTERIZATION OF SPATIAL MODELS
5
(A4) Heterogeneity: For every three distinct candidates {i, j, k} ⊆ C there is y ∈ Rd
such that i Ây j Ây k, and for every four distinct candidates {i, j, k, l} ⊆ C the sets
{y ∈ Rd : i ∼y j ∼y k} and {y ∈ Rd : i ∼y j ∼y l} are not equal.
The first property is standard. The second implies that if a voter with ideal point y
strictly prefers candidate i over j then any voter with ideal point sufficiently close to y
also prefers i over j. (A3) states that the set of voters preferring candidate i over j is
convex. Finally, (A4) requires the population of voters to be sufficiently diverse in its
preferences. Namely, for any (strict) ranking of every three candidates there should be
a voter who ranks these candidates according to that given order; And for every three
candidates there should be a voter that is indifferent between them but strictly prefers
some given fourth candidate over the three. We note that if m = 2 then (A4) is trivially
satisfied, and if m = 3 then the second part of (A4) is trivially satisfied.
Before stating our result we need one more definition.
Definition 1. Let {x1 , x2 , . . . , xm } ⊆ Rd and {v1 , v2 , . . . , vm } ⊆ R. We say that the
set {(x1 , v1 ), (x2 , v2 ), . . . , (xm , vm )} ⊆ Rd+1 is in a general position if the following two
conditions hold:
(i) For every distinct 1 ≤ i, j, k ≤ m, the vectors xi , xj , xk are affinely independent in Rd
(equivalently, xj − xi and xk − xi are linearly independent in Rd ).
(ii) For every distinct 1 ≤ i, j, k, l ≤ m, the set {y ∈ Rd : vi −||xi −y||2 = vj −||xj −y||2 =
vk − ||xk − y||2 = vl − ||xl − y||2 } is of dimension d − 3 at most5.
Informally speaking, if a set of points is not in a general position then it has a ‘degenerate structure’. We remark that if the points {(x1 , v1 ), (x2 , v2 ), . . . , (xm , vm )} are
independently drawn from some continuous distribution over Rd+1 then the resulting set
will be in a general position with probability 1. The precise meaning of the term general
position varies with the context in which it is used. The reader is referred to Matoušek
(2002, pp. 3-5), where this concept is discussed in greater detail.
Theorem 1. The following two statements are equivalent:
(i) The collection of binary relations {ºy }y∈Rd satisfies properties (A1) through (A4).
5The
dimension of a set A ⊆ Rd is defined as the dimension of the affine hull of A. In our case, the
set under consideration is the intersection of hyperplanes in Rd so it is an affine subspace. See Section
4 for details.
6
YARON AZRIELI
(ii) There are points {x1 , x2 , . . . , xm } ⊆ Rd and numbers {v1 , v2 , . . . , vm } ⊆ R such that
{(x1 , v1 ), (x2 , v2 ), . . . , (xm , vm )} is in a general position and, for every i, j ∈ C and every
y ∈ Rd , i ºy j if and only if vi − ||xi − y||2 ≥ vj − ||xj − y||2 .
The point xi is interpreted as the platform of candidate i, and vi is the score of i on
the valence dimension (1 ≤ i ≤ m). Theorem 1 states that properties (A1) through (A4)
are equivalent to the existence of {(x1 , v1 ), (x2 , v2 ), . . . , (xm , vm )} in a general position
such that each voter y ranks the candidates according to their score vi − ||xi − y||2 . Note
that all the voters agree on the location of the candidates in the policy space and on the
their score in the valence dimension.
3. Discussion
3.1. Uniqueness. Examining the proof of Theorem 1, one can see that the platforms
and valences derived from the properties (A1)-(A4) are not unique. However, we do have
the following connection between any two representations of the voters’ preferences.
Proposition 1. Assume {(x1 , v1 ), (x2 , v2 ), . . . , (xm , vm )} ⊆ Rd+1 represent the prefer0
ences {ºy }y∈Rd as in Theorem 1. Then {(x01 , v10 ), (x02 , v20 ), . . . , (x0m , vm
)} ⊆ Rd+1 also
represent {ºy }y∈Rd if and only if there is a positive number α > 0 and a vector β ∈ Rd
such that x0i = αxi + β for every 1 ≤ i ≤ m, and such that the equation6
1 0
(v − vj0 ) = vi − vj + (1 − α)(||xj ||2 − ||xi ||2 ) + 2β · (xi − xj )
α i
holds for every i, j ∈ C. In particular, if xi = x0i for 1 ≤ i ≤ m (i.e., α = 1 and β = 0)
(1)
then there is some γ ∈ R such that vi0 = vi + γ for 1 ≤ i ≤ m.
This result can be interpreted as follows. We may rescale and change the origin of
the policy space to get different sets of platforms that induce the same preferences.
But once the unit of measurement and the origin are fixed the platforms are uniquely
determined by the preferences. Moreover, once platforms are fixed, the relative valences
of the various candidates (the differences vi − vj ) are also unique.
3.2. Finite set of voters. A shortcoming of our model is that we assume that the
preferences of a voter with ideal point y are observable for every y ∈ Rd . It is much
more reasonable that the preferences of only a finite number of voters are observable. It
is tempting to try to get a similar result to that of Theorem 1 with a finite set of voters.
6For
two vectors z, w ∈ Rd we denote by z · w =
Pd
i=1 zi wi
the standard inner product in Rd .
CHARACTERIZATION OF SPATIAL MODELS
7
A natural modification of property (A3) in this case is that, for any i, j ∈ C, the convex
hull of the set of voters preferring candidate i over j and the convex hull of the set of
voters preferring candidate j over i are disjoint7.
However, we cannot get an analogue of Theorem 1 in this case. We demonstrate the
problem with the following example. Let d = 2, C = {1, 2, 3} and fix some ² > 0. The
set of voters, denoted Y , consists of six voters with the ideal points
Y = {y1 = (², ²), y2 = (−², −²), y3 = (−², −4), y4 = (², −4), y5 = (4, ²), y6 = (4, −²)}.
The preferences of these six voters are as follows. Voters {y2 , y3 } prefer candidate 1
over candidate 2 (the rest of the voters prefer candidate 2 over candidate 1). Voters
{y1 , y2 , y3 , y5 } prefer candidate 1 over candidate 3, and voters {y1 , y5 } prefer candidate
2 over candidate 3. Figure 1 illustrates the location of the voters’ ideal points in the
policy space and their rankings.
It is easy to check that the above condition of disjointness of the convex hulls is
satisfied. However, we claim that it cannot be the case that the voters agree on the
platforms and valences of the three candidates. Indeed, assume to the contrary that
there are {(x1 , v1 ), (x2 , v2 ), (x3 , v3 )} that represent these preferences as in Theorem 1.
The locations of the points y1 , y2 , y3 , y4 and the preferences of these voters imply that
the line {y ∈ R2 : v1 − ||y − x1 ||2 = v2 − ||y − x2 ||2 } should be close to both points (0, 0)
and (0, −4). Similarly, the line {y ∈ R2 : v1 −||y −x1 ||2 = v3 −||y −x3 ||2 } should be close
to both points (4, 0) and (0, −4), and the line {y ∈ R2 : v2 − ||y − x2 ||2 = v3 − ||y − x3 ||2 }
should be close to both points (0, 0) and (4, 0). For sufficiently small ², it must be the
case that the point ȳ = (1, −1) is in the triangle generated by these three lines. It means
that at this point we must have
v1 − ||ȳ − x1 ||2 < v2 − ||ȳ − x2 ||2 < v3 − ||ȳ − x3 ||2 < v1 − ||ȳ − x1 ||2 ,
which is impossible. If there was a voter with ideal point ȳ and transitive preferences
over candidates this could not have been happening. The characterization in the case of
a finite voter’s set remains unresolved.
3.3. Euclidean preferences. Our model does not presume any specific kind of preferences of the voters over the policy space8. The primitive only consists of rankings of the
candidates by voters. So properties (A1)-(A4) imply that voters agree on the platforms
7Assume
8We
for simplicity that only strict preferences are allowed.
thank Jim Peck for this remark.
8
YARON AZRIELI
and valences of the candidates and that the preferences of voter y over candidates can be
represented by uy (i) = vi − ||xi − y||2 . It is possible to assume from the start that voters
have Euclidean preferences. However, since utility is an ordinal notion, such modification
does not affect our results whatsoever.
The Euclidean norm is intimately related to convexity. Other norms typically induce
non-convex sets. It would be interesting to study preferences induced by other norms
than the Euclidean.
3.4. The valence dimension. Theorem 1 fails if we require all candidates to have the
same score (zero, w.l.o.g.) on the valence dimension9. Thus, more restrictions must be
imposed on voters’ preferences in order to allow a representation in the form ||xi − y||2 .
Finding natural additional properties that distinguish this case from the more general
one studied in this paper is an important direction for future research.
3.5. The cases m = 2 and m = 3. In contrast to the claim of subsection 3.4, if there
are only two or three candidates then it is possible to represent the voters’ preferences
without resorting to valences. The case m = 2 is trivial since one only needs to choose
the platforms x1 and x2 in equal distance from the hyperplane separating the voters that
prefer candidate 1 from those preferring candidate 2. In the case m = 3 we state this
fact as a proposition.
Proposition 2. Assume m = 3. The preferences {ºy }y∈Rd satisfy properties (A1)
through (A4) if and only if there are x1 , x2 , x3 ∈ Rd in a general position10 such that
i ºy j if and only if ||xi − y||2 ≤ ||xj − y||2 .
3.6. Observing just the first best. Theorem 1 crucially depends on property (A1).
In particular this means that we must observe the entire ranking of each voter over C.
In reality it is usually hard to extract this information from voters. A more plausible
assumption is that only the most preferred candidate(s) is (are) observed for each voter.
A possible way to formalize this is to assume that the primitive is a function f : Rd →
2C with the interpretation that f (y) ⊆ C is the set of candidates which voter y prefers
the most. We do not know how to get a similar result to that of Theorem 1 in this case
when the dimension of the policy space is d ≥ 2. However, it turns out that when d = 1
a simple characterization is possible (see the next subsection).
9The
reader is referred to Ash and Bolker (1985, Corollary 10) for a counter example.
10Here
we simply mean that these three vectors are affinely independent.
CHARACTERIZATION OF SPATIAL MODELS
9
3.7. The case d = 1. If the policy space is one dimensional (as is the case in many
papers) then Theorem 1 is no longer true, even if appropriately modified. The reason for
this failure is that the set of voters who are indifferent between some three candidates is
typically empty. This set plays a major role in the proof of the main result. Nevertheless,
we can get a representation similar to that of Theorem 1 if we assume that only the most
preferred candidates are observed for each voter (as in the previous subsection). We will
use the following properties for the characterization.
(B1) For every i ∈ C, the set {y ∈ R : f (y) = {i}} is not empty and open.
(B2) For every i ∈ C and y, z ∈ R, if i ∈ f (y) ({i} = f (y)) and i ∈ f (z) then
i ∈ f (αy + (1 − α)z) ({i} = f (αy + (1 − α)z)) for every α ∈ (0, 1).
Before stating the result, we need a definition analogue to Definition 1.
Definition 2. The set {(x1 , v1 ), (x2 , v2 ), . . . , (xm , vm )} ⊆ R2 is well-ordered if there is a
permutation π : C → C such that the following two conditions hold:
(i) xπ(1) < xπ(2) < . . . < xπ(m) .
(ii) aπ(1)π(2) < aπ(2)π(3) < . . . < aπ(m−1)π(m) where aπ(i)π(i+1) =
x2π(i) −x2π(i+1) +vπ(i+1) −vπ (i)
2(xπ(i) −xπ(i+1) )
for
i = 1, 2, . . . , m − 1.
Proposition 3. The correspondence f : R → 2C satisfies properties (B1) and (B2) if
and only if there is a well-ordered set {(x1 , v1 ), (x2 , v2 ), . . . , (xm , vm )} ⊆ R2 such that
f (y) = argmax{vi − (xi − y)2 : i ∈ C}.
4. Proofs
4.1. Proof of Theorem 1. The proof of Theorem 1 is similar to the proof of the main
result in Azrieli and Lehrer (2007). We therefore provide an outline of the proof and
only detail those steps that did not appear in that paper.
A simple but important observation is that, for any xi 6= xj ∈ Rd and vi , vj ∈ R, the
set {y ∈ Rd : vi − ||xi − y||2 = vj − ||xj − y||2 } is an affine subspace of dimension d − 1 (a
hyperplane), perpendicular to the direction xi − xj . Indeed, a simple computation shows
that this set can be rewritten as {y ∈ Rd : y · (xi − xj ) = 12 (vj − vi + ||xi ||2 − ||xj ||2 )}.
Similarly, the set {y ∈ Rd : vi − ||xi − y||2 > vj − ||xj − y||2 } is an open half space in
Rd (given that xi 6= xj ).
(ii) implies (i):
10
YARON AZRIELI
Fix the sets {x1 , x2 , . . . , xm } ⊆ Rd and {v1 , v2 , . . . , vm } ⊆ R. Property (A1) is obviously satisfied. Denote Aij = {y ∈ Rd : vi − ||xi − y||2 = vj − ||xj − y||2 } and
Bij = {y ∈ Rd : vi − ||xi − y||2 > vj − ||xj − y||2 }. By property (i) of Definition 1,
xi 6= xj for every i 6= j ∈ C. Thus, each Bij is open and convex and each Aij is the
boundary of the closed half space Bij ∪ Aij . This shows that properties (A2) and (A3)
are satisfied.
Property (A4) is satisfied because the set {(x1 , v1 ), (x2 , v2 ), . . . , (xm , vm )} is in a general
position. Indeed, take any distinct i, j, k ∈ C. We need to show that there is some y
with i Ây j Ây k. If this was not true then it must be that Bij and Bjk do not intersect.
But this can only happen if xi − xj and xj − xk are linearly dependent, a contradiction to
the assumption of general position (property (i)). Finally, take any distinct i, j, k, l ∈ C.
We need to show that Bij ∩ Bjk and Bij ∩ Bjl are not equal. The set Bij ∩ Bjk is the
intersection of two (non-parallel) hyperplanes in Rd and so it is of dimension d − 2. If
Bij ∩ Bjk = Bij ∩ Bjl then Bij ∩ Bjk ∩ Bjl = Bij ∩ Bjk is also of dimension d − 2,
contradicting the assumption of general position (property (ii)).
¤
(i) implies (ii):
The proof is constructive. We first find the platforms x1 , x2 , . . . , xm of the candidates, and then construct the valences v1 , v2 , . . . , vm . We need however to state some
preliminary claims. The proofs of all these claims can be found in Azrieli and Lehrer
(2007).
Claim 1. For every ordered pair (i, j) of distinct candidates there is a non-zero vector
sij ∈ Rd and a number cij ∈ R such that {y ∈ Rd : i ºy j} = {y ∈ Rd : sij · y ≤ cij }.
Moreover, these vectors and numbers can be chosen such that sji = −sij and cji = −cij
for every (i, j).
Fix a collection {sij , cij }i,j∈C as in Claim 1 until the end of the proof.
Claim 2. (A4) implies that, for every i, j, k ∈ C, the vectors sij and sik are linearly
independent.
Claim 3. For every i, j, k ∈ C, the vectors sij , sik and sjk are not linearly independent.
For t, s ∈ Rd , denote by R(t, s) the ray that starts at t and continues in the direction of
s. That is R(t, s) = {t + αs : α ≥ 0}.
Claim 4. If x1 , x2 ∈ Rd satisfy x2 −x1 = αs12 for some α > 0 then, for every 3 ≤ i ≤ m,
the rays R(x1 , s1i ) and R(x2 , s2i ) intersect .
CHARACTERIZATION OF SPATIAL MODELS
11
We are now in the position to construct the sets {x1 , x2 , . . . , xm } and {v1 , v2 , . . . , vm }.
The point x1 is chosen arbitrarily. Next, define x2 = x1 + α12 s12 , where α12 > 0 is
arbitrary. For every 3 ≤ i ≤ m, define xi to be the unique point of intersection (by
Claim 4) of the rays R(x1 , s1i ) and R(x2 , s2i ). A key point in the proof is that, when
{x1 , x2 , . . . , xm } are defined in this way, then, for every 1 ≤ i, j ≤ m, xj − xi = αij sij
for some αij > 0. This fact follows from Proposition 1 (page 26) in Azrieli and Lehrer
(2007). Finally, choose v1 arbitrarily and define vi = v1 − ||x1 ||2 + ||xi ||2 − 2α1i c1i for
every 2 ≤ i ≤ m.
It is useful to note that αij sij = α1j s1j − α1i s1i for every 3 ≤ i, j ≤ m. Indeed, the
left–hand side of the equality is xj − xi while the right–hand side is (xj − x1 ) − (xi − x1 ).
This implies also that αij cij = α1j c1j − α1i c1i . To see this, take y ∈ Rd such that 1 ∼y i
and 1 ∼y j (the existence of such y is guaranteed by Claim 2). Transitivity implies that
i ∼y j. So y · s1i = c1i , y · s1j = c1j and y · sij = cij . Multiplying these equalities by
α1i , α1j and αij correspondingly, and subtracting the first from the second we get the
above equality.
To complete the proof we need to check that the set {(x1 , v1 ), (x2 , v2 ), . . . , (xm , vm )}
is in a general position and that, for every i, j ∈ C and y ∈ Rd , i ºy j if and only if
vi − ||xi − y||2 ≥ vj − ||xj − y||2 . For the latter we have
i ºy j ⇐⇒ sij · y ≤ cij ⇐⇒ (xj − xi ) · y ≤ αij cij ⇐⇒ (xj − xi ) · y ≤ α1j c1j − α1i c1i
¢ 1¡
¢
1¡
⇐⇒ (xj − xi ) · y ≤
v1 − vj + ||xj ||2 − ||x1 ||2 −
v1 − vi + ||xi ||2 − ||x1 ||2
2
2
¢
1¡
⇐⇒ (xj − xi ) · y ≤
vi − vj + ||xj ||2 − ||xi ||2 ⇐⇒ vi − ||xi − y||2 ≥ vj − ||xj − y||2 .
2
For the former, the vectors xi , xj , xk are affinely independent since xj − xi = αij sij and
xk − xi = αik sik , and these are linearly independent vectors by Claim 2. Finally, each of
the sets {y ∈ Rd : i ∼y j ∼y k} and {y ∈ Rd : i ∼y j ∼y l} is an affine subspace of
dimension d − 2. By (A4) they are not equal so their intersection is of dimension d − 3
at most. This proves that {(x1 , v1 ), (x2 , v2 ), . . . , (xm , vm )} is in a general position.
¤
4.2. Proof of Proposition 1. First, it is easy to check that if there are α > 0 and
β ∈ Rd such that x0i = αxi + β for 1 ≤ i ≤ m, and in addition equation (1) is satisfied
0
)} represent the
then {(x1 , v1 ), (x2 , v2 ), . . . , (xm , vm )} and {(x01 , v10 ), (x02 , v20 ), . . . , (x0m , vm
same preferences.
0
)}
Now, assume that {(x1 , v1 ), (x2 , v2 ), . . . , (xm , vm )} and {(x01 , v10 ), (x02 , v20 ), . . . , (x0m , vm
represent the same preferences {ºy }y∈Rd . It follows from the proof of Theorem 1 that for
12
YARON AZRIELI
every i, j ∈ C there is a positive number, say tij > 0, such that xj −xi = tij (x0j −x0i ) (with
the convention tij = −tji ). Fix some three candidates i, j, k ∈ C. Sum up the equalities
xj − xi = tij (x0j − x0i ), xi − xk = tki (x0i − x0k ), xk − xj = tjk (x0k − x0j ) and rearrange the
terms to obtain (x0i − x0j )(tki − tij ) + (x0k − x0j )(tjk − tki ) = 0. But the vectors x0i , x0j , x0k are
affinely independent so tki − tij = tjk − tki = 0. It follows that tij = tki = tjk , so there is a
number α > 0 such that xj −xi = α(x0j −x0i ) for every i, j ∈ C. Now, define β = x1 −αx01 .
For every 2 ≤ i ≤ m we have x1 − xi = α(x01 − x0i ) or xi − αx0i = x1 − αx01 = β. That is,
x0i = αxi + β for every 1 ≤ i ≤ m.
Finally, we must have
1
2
(vi − vj + ||xj ||2 − ||xi ||2 ) =
1
2
¡
¢
vi0 − vj0 + ||x0j ||2 − ||x0i ||2 for
every i, j ∈ C. Substituting αxi + β for x0i and αxj + β for x0j and rearranging we obtain
equation (1). In particular, if x0i = xi and x0j = xj then vi0 − vi = vj0 − vj . Define
γ = v10 − v1 . It follows that vi0 = vi + γ for every 1 ≤ i ≤ m.
¤
4.3. Proof of Proposition 2. The if part follows from Theorem 1, so we only need
to prove the only if part. By Theorem 1, there are (x1 , v1 ), (x2 , v2 ), (x3 , v3 ) in a general
position that represent the preferences. It follows that the vectors x1 − x2 and x1 − x3
are linearly independent. Therefore, there is β ∈ Rd that solves the two equations
β · (x1 − x2 ) =
v2 −v1
2
and β · (x1 − x3 ) =
v3 −v1
.
2
Notice that the same vector β must satisfy
v3 −v2
.
2
also β · (x2 − x3 ) =
Define x0i = xi + β for i = 1, 2, 3.
By Proposition 1, the set {(x01 , v10 ), (x02 , v20 ), (x03 , v30 )} represent the same preferences as
{(x1 , v1 ), (x2 , v2 ), (x3 , v3 )} if the equation vi0 − vj0 = vi − vj + 2β · (xi − xj ) is satisfied for
v −v
every i, j ∈ C. By construction, the vector β satisfies β · (xi − xj ) = j 2 i for every i, j. It
follows that v10 = v20 = v30 = 0 solve the above equations. That is, {(x01 , 0), (x02 , 0), (x03 , 0)}
represent the preferences {ºy }y∈Rd .
¤
4.4. Proof of Proposition 3. Assume first that the correspondence f can be represented as in the proposition. We can assume w.l.o.g. that π is the identity, so x1 < x2 <
. . . < xm and a12 < a23 < . . . < a(m−1)m . It is also convenient to denote a01 = −∞
and am(m+1) = +∞. Now, for every 1 ≤ i ≤ m − 1, a simple computation shows that
vi − (xi − y)2 ≥ vi+1 − (xi+1 − y)2 if and only if y ≤ ai(i+1) (the same equivalence
holds when the weak inequalities are replaced by strict ones). It follows that candidate
i (1 ≤ i ≤ m) is the unique maximizer of {vj − (xj − y)2 : j ∈ C} if and only if
y ∈ (a(i−1)i , ai(i+1) ) and that i is a maximizer (not necessarily unique) of this expression
if and only if y ∈ [a(i−1)i , ai(i+1) ]. This shows that f satisfies properties (B1) and (B2).
CHARACTERIZATION OF SPATIAL MODELS
13
Conversely, assume that f satisfies (B1) and (B2). These properties imply that there
is a permutation of the candidates, w.l.o.g. the identity, and a sequence of numbers
a12 < a23 < . . . < a(m−1)m such that f (y) = {i} if and only if y ∈ (a(i−1)i , ai(i+1) ) and
f (y) = {i, i + 1} if and only if y = ai(i+1) for 1 ≤ i ≤ m.
Take any set of points x1 < x2 < . . . < xm . Define v1 = 0 and, for every 1 ≤ i ≤ m−1,
let vi+1 = 2ai(i+1) (xi −xi+1 )−x2i +x2i+1 +vi . Rearranging, this gives ai(i+1) =
x2i −x2i+1 +vi+1 −vi
2(xi −xi+1 )
for i = 1, 2, . . . , m−1. Thus, the set {(x1 , v1 ), (x2 , v2 ), . . . , (xm , vm )} ⊆ R2 is well-ordered.
Finally, we need to check that f (y) = argmax{vi − (xi − y)2 : i ∈ C}. This is true since
i ∈ f (y) ⇐⇒ y ∈ [a(i−1)i , ai(i+1) ] ⇐⇒
x2i−1 − x2i + vi − vi−1
x2 − x2i+1 + vi+1 − vi
≤y≤ i
2(xi − xi+1 )
2(xi−1 − xi )
⇐⇒ vi − (y − xi )2 ≥ vi−1 − (y − xi−1 )2 and vi − (y − xi )2 ≥ vi+1 − (y − xi+1 )2
⇐⇒ vi − (y − xi )2 ≥ vj − (y − xj )2 for all j 6= i.
¤
References
[1] Ansolabehere, S. and J. M. Snyder, JR (2000) Valence politics and equilibrium in spatial election
models, Public Choice 51, 327-336.
[2] Aragones, E. and T. R. Palfrey (2002) Mixed equilibrium in a Downasian model with a favored
candidate, Journal of Economic Theory 103, 131-161.
[3] Ash, P. F. and E. Bolker (1985) Recognizing Dirichlet tessellation, Geometria Dedicata 19, 175-206.
[4] Ash, P. F. and E. Bolker (1986) Generalized Dirichlet tessellation, Geometria Dedicata 20, 209-243.
[5] Aurenhammer, F. (1987) Power diagrams: Properties, algorithms and applications, SIAM Journal
on Computing 16, 78-96.
[6] Azrieli, Y. and E. Lehrer (2007) Categorization generated by extended prototypes – an axiomatic
approach, Journal of Mathematical Psychology 51, 14-28.
[7] Boots, B., A. Okabe and K. Sugihara (1992) Spatial Tessellations, John Wiley and Sons Ltd,
England.
[8] Degan, A. (2007) Candidate valence: Evidence from consecutive presidential elections, International
Economic Review 48, 457-482.
[9] Dix, M. and R. Santore (2002) Candidate ability and platform choice, Economics letters 76, 189194.
[10] Downs, A. (1957) An Economic Theory of Democracy, Harper, New York.
[11] Enelow, J. and M. J. Hinich (1981) A new approach to voter uncertainty in the Downasian spatial
model, American Journal of Political Science 25, 483-493.
[12] Gersbach, H. (1998) Communication skills and competition for donors, European Journal of Political
Economy 14, 3-18.
14
YARON AZRIELI
[13] Gilboa, I. and D. Schmeidler (2001) A Theory of Case-Based Decisions, Cambridge University
Press.
[14] Groseclose, T. (2001) A model of candidate location when one candidate has a valence advantage,
American Journal of Political Science 45, 862-886.
[15] Hotelling, H. (1929) Stability in competition, Economic Journal 39, 41-57.
[16] Kim, K. (2005) Valence characteristics and entry of a third party, Economics Bulletin 4, 1-9.
[17] Matoušek, J. (2002) Lectures on Discrete Geometry, Springer-Verlag, New York, USA.
[18] Myerson, R. B. (1995) Axiomatic derivation of scoring rules without the ordering assumption, Social
Choice and Welfare 12, 59-74.
[19] Schofield, N. (2007) The mean voter theorem: Necessary and sufficient conditions for convergent
equilibrium, Review of Economic Studies 74, 965-980.
[20] Smith, J. (1973) Aggregation of preferences with a variable electorate, Econometrica 41, 1027-1041.
[21] Young, H. P. (1975) Social choice scoring functions, Siam Journal on Applied Mathematics 28,
824-838.
y5
y1
2
1
2
3
1 3
y2
1
3 2
3
(1,-1)
y4
y3
1
3 2
3
2 1
Figure 1
y6
2 1